MegaAgent: A Large-Scale Autonomous LLM-based Multi-Agent System Without Predefined SOPs

Abstract: LLM-based multi-agent systems (MAS) have shown promise in tackling complex tasks. However, existing solutions often suffer from limited agent coordination and heavy reliance on predefined Standard Operating Procedures (SOPs), which demand extensive human input. To address these limitations, we propose MegaAgent, a large-scale autonomous LLM-based multi-agent system. MegaAgent generates agents based on task complexity and enables dynamic task decomposition, parallel execution, efficient communication, and comprehensive system monitoring of agents. In evaluations, MegaAgent demonstrates exceptional performance, successfully developing a Gobang game within 800 seconds and scaling up to 590 agents in a national policy simulation to generate multi-domain policies. It significantly outperforms existing systems, such as MetaGPT, in both task completion efficiency and scalability. By eliminating the need for predefined SOPs, MegaAgent demonstrates exceptional scalability and autonomy, setting a foundation for advancing true autonomy in MAS. Our code is available at https://github.com/Xtra-Computing/MegaAgent .

• The paper demonstrates a novel hierarchical framework that dynamically splits tasks for autonomous agent generation and coordination in LLM-based systems.
• It utilizes multi-level task splitting, parallel execution, and integrated modules (action, storage, monitor) to optimize performance, with empirical validation via Gobang and policy simulation experiments.
• The framework significantly enhances scalability and efficiency by enabling dynamic, concurrent operations without relying on predefined SOPs, paving the way for future agent systems.

MegaAgent: A Large-Scale Autonomous LLM-Based Multi-Agent System

Introduction

MegaAgent is an innovative framework designed to address the limitations of predefined Standard Operating Procedures (SOPs) and scalability in LLM-powered multi-agent systems (LLM-MA systems). Traditional systems often lack autonomy and rely on fixed protocols, leading to inefficiencies in complex task environments. MegaAgent introduces a hierarchical structure to dynamically generate agents based on task requirements, enabling effective cooperation and concurrent operations without predefined SOPs. This approach enhances scalability, autonomy, and performance, paving the way for future developments in large-scale autonomous systems.

Framework Structure

MegaAgent comprises several key components: Multi-Level Task Splitting, Action Module, Storage Module, Monitor Mechanism, and Communication Mechanism.

1. Multi-Level Task Splitting: MegaAgent starts by dividing a given task into multiple hierarchical sub-tasks, each assigned to a group of agents. A boss agent orchestrates task assignment, ensuring that tasks are properly distributed among sub-task admins who oversee individual agent groups. This hierarchical design facilitates task management and scalability by distributing complexity across different levels, allowing for efficient handling of large-scale interactions.
2. Action Module: This component is responsible for executing sub-tasks. Agents within a group create checklists to manage their actions, while an OS agent monitors outputs to ensure they meet predefined formats. The Action Module supports parallelism, allowing multiple agent groups to operate simultaneously, significantly reducing task completion time compared to sequential systems.
3. Storage Module: It facilitates communication between agents and external files, integrating Git management, memory databases for long-term storage, and a task monitor to ensure task progress and organization. This module helps maintain data consistency and provides a comprehensive framework for collaborative agent operations.
4. Monitor Mechanism: This employs an OS agent and checklists to minimize errors and ensure that agents follow expected workflows. Each agent tracks its progress through checklists, which are regularly updated to ensure alignment with overarching goals.
5. Communication Mechanism: Agents communicate within defined hierarchies, interacting with direct superiors or subordinates. This structured format supports effective collaboration within and across agent groups, enhancing overall communication efficiency.

   Figure 1: MegaAgent starts by accepting prompts, which it automatically divides into tasks for multiple agent groups. OS Agent monitors the agents' progress, while the action module interacts with the storage module through a data retriever.

Experimental Evaluation

MegaAgent was evaluated through two experiments: Gobang game development and national policy simulation.

1. Gobang Game Development: MegaAgent demonstrated superior autonomy and efficiency, completing the task within 800 seconds with a naive AI, outperforming other LLM-MA systems like AutoGen, MetaGPT, CAMEL, and AgentVerse. The framework successfully coordinated task execution across seven agents, leveraging parallelism to optimize performance.

   Figure 2: Interface of MegaAgent's Gobang Demo

2. National Policy Simulation: MegaAgent showcased its scalability by managing approximately 590 agents to generate detailed policies across various sectors within 3000 seconds. The framework's hierarchical structure facilitated efficient policy development, highlighting its capability to handle complex, large-scale scenarios.

   Figure 3: An example of the generated hierarchy for the policy making experiment

Implications and Future Directions

MegaAgent establishes a foundation for autonomous cooperation in LLM-MA systems, serving as a potential operating system for future large-scale agent frameworks. It addresses the critical need for scalable infrastructure and dynamic task management in complex environments. Future research may explore optimizing communication efficiency, integrating diverse LLMs for cost-effective operations, and further reducing hallucinations by leveraging external knowledge bases.

Conclusion

MegaAgent presents a practical solution for managing large-scale LLM-MA systems, emphasizing autonomy, scalability, and efficient cooperation. Through innovative hierarchical structuring and parallel execution capabilities, it surpasses existing frameworks in task completion efficacy. The framework lays the groundwork for future exploration in autonomous multi-agent systems, drawing attention to enhancing effective agent cooperation in increasingly complex scenarios.